Phase 0B — Observatory + evaluation baseline #2

Workbook: phases/00b-observatory/
Built in: agent-observatory + benchmarks

Gate passed. Instrument still produced five void experiments — the gap between those two statements is the real lesson of this phase.

Labs

• 0B.1 — Benchmark before dashboard
• 0B.2 — 5 plain-agent repetitions
• 0B.3 — Collector with debug exporter only
• 0B.4 — Tempo + Grafana, read one trace
• 0B.5 — B0 variance experiment

Exit gate

• Deterministic benchmark works
• Hidden tests catch an intentionally bad implementation
• 5 baseline runs exist
• One trace visible in Grafana
• Run records normalized independently of vendor OTel shape

Still open upstream

• agent-observatory Reviewing a contract before it leaves the machine is now the default #41 — OTel privacy + cardinality at the collector boundary
• agent-observatory Phase 9 — Memory #11 — bank privacy checklist
• BehaviorDto counters non-nullable, so a telemetry gap is indistinguishable from a genuine zero